Mixed model multivariate time sequence anomaly detection method based on graph neural network

Abstract

The invention discloses a mixed model multivariate time sequence anomaly detection method based on a graph neural network, and the method comprises the steps: dividing a multivariate time sequence into a feature matrix based on a sliding window, an adjacent matrix, and an adjacent matrix based on a fixed window, and carrying out the preprocessing of a first feature matrix, a first adjacent matrix, and a second adjacent matrix; constructing a graph convolutional neural network prediction model, and inputting the first feature matrix and the first adjacent matrix to obtain a prediction value; comparing the real value with the abnormal time stamp to judge an abnormal time stamp; constructing a convolutional neural network and attention long-short-term memory network hybrid reconstruction model, and inputting the second adjacent matrix to obtain a reconstructed adjacent matrix; comparing to obtain a reconstruction error matrix, and judging an abnormal time sequence according to the sizes of the elements in the reconstruction error matrix and the number of the elements exceeding a threshold value; and determining an abnormal point according to the abnormal timestamp and the abnormal time sequence. Compared with the prior art, the abnormal time stamp and the abnormal time sequence in the multivariate time sequence can be detected, and the abnormal detection granularity, efficiency and detection accuracy of the multivariate time sequence are improved.

Description

technical field [0001] The present application relates to the technical field of multivariate time-series data detection, and more specifically, relates to a mixed model multivariate time-series anomaly detection method based on a graph neural network. Background technique [0002] Time series refers to a series of data collected at a certain time interval, and the data can be unary or multivariate. Multivariate time-series data widely exists in social industrial production processes. Detecting abnormal events in multivariate time-series is an effective method for performance reliability assurance, and it is also the basis for further fault prediction, abnormal location, rapid stop loss, and root cause analysis of faults. Multivariate time series anomaly detection is a critical step in building a safe and reliable system that can be timely to minimize the loss caused by anomalous events. The efficiency and accuracy of anomaly detection and analysis are of great significance...

AI Technical Summary

Problems solved by technology

[0003] Regarding the anomaly detection of multivariate time series, some existing technologies can detect the anomaly of the entire multivariate time series at a certain time stamp or time period, but cannot detect which multivariate time series is abnormal among the multivariate time series, that is, abnormal time series detection; Some existing technologies can detect which time series is abnormal among multiple time series, or which time series is abnormal among multiple time series in a certain period of time, but these cannot accurately detect which time stamp of the abnormal time series in the multivariate time series is abnormal , that is, abnormal timestamp detection

It can be seen that, for multivariate timing anomaly detection, the current existing technology cannot realize the detection of abnormal timing and the detection of abnormal time stamp at the same time

Images